Free ChatGPT web tier, dressed up as an OpenAI API

What it does

This Go project is a thin proxy that sits between your code and ChatGPT’s web backend, translating standard OpenAI API requests into whatever the ChatGPT web app expects. It supports both anonymous (no-login) web sessions and registered account tokens, then returns the responses in familiar stream or JSON format. In short, it lets you hit chatgpt.com through a /v1/chat/completions endpoint as if it were a real API.

The interesting bit

The real trick is the hardcoded model mapping: every gpt-3.5-turbo variant is translated to text-davinci-002-render-sha, the internal identifier for ChatGPT’s free web model. That sleight-of-hand lets client libraries talk to the web UI without knowing they’re not on a real API. The project also maintains a pool of ACCESS_TOKENs via a YAML file and exposes how many are currently usable.

Key highlights

• Supports anonymous, bearer-token, and ac- prefixed access-token authentication modes
• Exposes an OpenAI-compatible /v1/chat/completions endpoint with streaming and JSON responses
• Pools account tokens via a YAML file and exposes a /v1/accTokens availability check
• Hardcodes model aliases so gpt-3.5-turbo requests hit the web UI’s backend model
• Ships with one-click deploy buttons for Docker, Vercel, and Koyeb

Caveats

• The README warns that native IPs no longer work for no-login access; you now need an outbound proxy to reach the free tier
• All gpt-3.5-turbo variants map to the same web backend model, so model selection is largely cosmetic
• Documentation and configuration comments are entirely in Chinese, which may slow down non-Chinese-speaking operators

Verdict Worth a look if you want a self-hosted bridge to ChatGPT’s web tier without a paid API key; avoid it if you need official SLAs or English-language documentation.